training_code/data_config.yaml

# Data config: pusht — 100 Franka teleop demonstrations of the real-hardware
# Push-T task (HF dataset SleepMastger/pusht-manipulation), 32,131 frames at
# 10 Hz across 4 collection sessions. Converted to LeRobot v2.1 by
# lift2lerobot/convert_lift_hdf5_to_lerobot_v21.py (--fps 10) via the
# raw/pusht_layout symlink shim.
#
# Task string (deploy-critical, must byte-match at inference, NO trailing
# period): "push the T block to the target outline"
#
# DEGENERATE CHANNELS — read before changing norm settings.
# The teleop rig commanded translation only and the gripper stayed clamped on
# the pusher pen for the whole task, so 6 of the 15 numeric channels are
# constant across all 32,131 frames (measured, not assumed):
#     action  drx, dry, drz   range 0.0
#     action  gripper         range 0.0   (raw +1 -> (1-g)/2 -> 0.0)
#     state   gripL, gripR    range 3.2e-05
# The dataset README warns this makes a min/max normalizer divide by zero.
# It does NOT here: SingleFieldLinearNormalizer (normalizer.py:96-118) has
# `ignore_dim = input_range < range_tol` with range_tol = 1e-4, which catches
# all six (the tightest, 3.2e-05, clears it by 3x) and gives them scale 1.0
# with offset -min instead of dividing by ~0. So an ignored dim normalizes to
# (x - min): exactly 0 for the four action dims whose raw range is exactly
# 0.0, and [0, 3.2e-05] for the two gripper state dims. Either way the output
# is finite and bounded by the raw range, i.e. below range_tol — negligible
# next to the +/-1 live channels. Verified end to end by preflight.py, not
# just read off the source. Keeping the full 7-dim action / 8-dim proprio
# costs nothing and keeps the architecture byte-identical to the dish_utensil
# and fruit runs, so results stay comparable. If you ever raise range_tol
# below 3.2e-05, the gripper state dims WILL go inf — don't.
#
# Staged by stage_pusht_local.sh to node-local /tmp (NOT /dev/shm — 853 wipe).
train:
  _target_: fastwam.datasets.lerobot.robot_video_dataset.RobotVideoDataset
  dataset_dirs:
    - /tmp/george_pusht/pusht_lerobot_v21
  shape_meta:
    images:
      - key: image
        raw_shape: [3, 256, 256]
        shape: [3, 224, 224]
      - key: wrist_image
        raw_shape: [3, 256, 256]
        shape: [3, 224, 224]
    action:
      - key: default
        raw_shape: 7 # before transform
        shape: 7 # after transform
    state:
      - key: default
        raw_shape: 8 # before transform
        shape: 8 # after transform
  num_frames: 33
  global_sample_stride: 1
  action_video_freq_ratio: 4 # 32 action, 9 video frames
  video_size: [224, 448] # final resize video
  camera_key: null
  val_set_proportion: 0.0
  is_training_set: true
  skip_padding_as_possible: false
  concat_multi_camera: "horizontal"
  processor:
    _target_: fastwam.datasets.lerobot.processors.fastwam_processor.FastWAMProcessor
    shape_meta: ${data.train.shape_meta}
    num_obs_steps: ${data.train.num_frames}
    num_output_cameras: 2
    action_output_dim: 7 # eef_pose (6) + gripper (1)
    proprio_output_dim: 8 # eef_pose (6) + gripper (2)
    delta_action_dim_mask:
      # The raw HDF5 actions are ALREADY per-step deltas (dx,dy,dz in metres
      # per 100 ms), so this mask is not a "convert to delta" switch — the
      # processor only uses it to zero PADDED action steps on the dims where
      # 0 is the neutral value (fastwam_processor.py:300-308).
      default: [true, true, true, true, true, true, false] # eef poses are delta, gripper is not

    # action & state normalization
    action_state_transforms: null
    use_stepwise_action_norm: False
    norm_default_mode: min/max
    norm_exception_mode: null

    action_state_merger:
      _target_: fastwam.datasets.lerobot.transforms.action_state_merger.ConcatLeftAlign

    train_transforms:
      - _target_: fastwam.datasets.lerobot.transforms.image.ToTensor
      - _target_: torchvision.transforms.Resize
        size: [224, 224]
    val_transforms:
      - _target_: fastwam.datasets.lerobot.transforms.image.ToTensor
      - _target_: torchvision.transforms.Resize
        size: [224, 224]
  text_embedding_cache_dir: /tmp/george_pusht/pusht_text_embeds_cache
  context_len: 128